From: Potential applications of artificial intelligence in image analysis in cornea diseases: a review
Year | Authors | Imaging modality | Sample size (eyes) | Study population | Outcome measures | AI algorithms | Diagnostic performance | Validation model |
---|---|---|---|---|---|---|---|---|
Corneal endothelium | ||||||||
 2023 | Karmakar et al. [75] | Konan CellCheck XL | 612 | Healthy and diseased eyes | Segmentation of endothelial cells | Mobile-CellNet CNN | Mean absolute error: 4.06% | Hold-out validation |
 2022 | Qu et al. [136] | IVCM | 97 | Healthy, FECD and corneal endotheliitis eyes | Segmentation of endothelial cells | CNN | PCC: 0.818–0.932 | Hold-out validation |
 2020 | Canavesi et al. [77] | GDOCM | 10 | Eye bank | Segmentation of endothelial cells | CNN | Correlation: 0.91–0.94 | Cross validation |
 2019 | Bennett et al. [80] | JDS Uniphase, TOMEY TMS-5 | 10 | Healthy eyes | Evaluation of corneal thickness | CNN | RMSE: 0.045–0.048 Acc: 84.82%–89.26% | Hold-out validation |
 2019 | Vigueras-Guillén et al. [137] | Topcon SP-1P | 738 | Patients with Baerveldt glaucoma device and DSAEK | Segmentation of endothelial cells | CNN | Mean absolute error: 4.32%–11.74% | Hold-out validation |
 2019 | Daniel et al. [70] | Topcon SP-3000 | 385 | Database of healthy, endothelial disease and corneal graft eyes | Segmentation of endothelial cells | U-Net CNN | PCC: 0.96, Sens: 0.34% Precis: 0.84% | Hold-out validation |
 2018 | Fabijańska et al. [73] | Specular microscopy | 30 | Dataset of endothelial cell images | Evaluation of corneal thickness | U-Net CNN | AUC: 0.92, Dice: 0.86 Mean absolute error: 4.5% | Hold-out validation |
 2018 | Vigueras-Guillén et al. [76] | Topcon SP-1P | 103 | Dataset of endothelial cell images | Evaluation of corneal thickness | SVM | Precis: P < 0.001 Acc: P < 0.001 | Cross validation |
Corneal nerves | ||||||||
 2023 | Li et al. [93] | HRT-3 confocal microscopy | 30 | Eyes with slight xerophthalmia | Reconstruction of CSNP in images | NerveStitcher CNN | No validation or qualitative evaluation | N.A |
 2022 | Setu et al. [88] | IVCM | 197 | Healthy and DED eyes | Segmentation of CNF and DC | U-Net, Mask R CNNs | Sens: 86.1%–94.4%, Spec: 90.1% Precis: 89.4%, ICC: 0.85–0.95 | Cross validation |
 2022 | Mou et al. [89] | HRT-3 confocal microscopy | 300 | CORN1500 dataset images | Grading of corneal nerve tortuosity | ImageNet, AuxNet | Acc: 85.64% | Cross validation |
 2021 | Zéboulon et al. [95] | AS-OCT | 607 | Healthy and edematous corneas | Measurement of edema fraction | CNN | Threshold for diagnosis: 6.8%, AUC: 0.994, Acc: 98.7% Sens: 96.4%, Spec: 100% | Hold-out validation |
 2021 | Deshmukh et al. [96] | ASP | 504 | Genetically confirmed GCD2 patients | Segmentation of cornea lesions | U-Net, CNN | IoU: 0.81 Acc: 99% | Cross validation |
 2021 | Salahouddin et al. [138] | CCM | 534 | Healthy and type I diabetic eyes | DPN detection | U-net CNN | κ: 0.86, AUC: 0.86–0.95 Sens: 84%–92%, Spec: 71%–80% | Hold-out validation |
 2021 | McCarron et al. [86] | HRT-3 confocal microscopy | 73 | Healthy and SIV-infected macaque eyes | Characterize difference in CSNP in acute SIV infection | deepNerve CNN | SIV infection reduced CNFL and fractal dimension (P = 0.01, P = 0.008) | N.A |
 2021 | Yıldız et al. [139] | HRT-3 confocal microscopy | 85 | Healthy and chronic ocular surface pathology eyes | Segmentation of CSNP | GAN, U-Net CNN | PCC: 0.847–0.883 AUC: 0.8934–0.9439 | N.A |
 2020 | Scarpa et al. [85] | CCM | 100 | Healthy and DPN eyes | Classification of DPN and healthy eyes | CNN | Acc: 96% | Cross validation |
 2020 | Williams et al. [84] | CCM | 2137 | Healthy and DPN eyes | Quantification of CSNP, detection of DPN | CNN | ICC: 0.656–0.933, AUC: 0.83 Spec: 87%, Sens: 68% | Hold-out validation |
 2020 | Wei et al. [140] | HRT-3 confocal microscopy | 139 | Healthy eyes | Segmentation of CSNP | CNS-Net CNN | AUC: 0.96, Precis: 94% Sens: 96%, Spec: 75% | Hold-out validation |